My iPhone 16 Pro Max produces garbage output when running MLX LLMs (journal.rafaelcosta.me)

Article summary

The author's iPhone 16 Pro Max produces incorrect output when running MLX LLMs, while an iPhone 15 Pro and a MacBook Pro run the same code perfectly. The issue is suspected to be a hardware defect in the Neural Engine or another ML-related system. The author debugged the issue by comparing tensor outputs on different devices and found that the iPhone 16 Pro Max's output was orders of magnitude off. The problem was resolved when the author upgraded to an iPhone 17 Pro Max.

Main themes

• iPhone hardware issues
• Machine Learning
• Neural Engine
• Debugging
• Apple devices

What commenters say

• The article's debugging process is a great example of how to approach complex technical issues.
• The failure of the iPhone 16 Pro Max to run MLX LLMs correctly is likely a hardware defect rather than a software issue.
• Some commenters prefer to use third-party calculator apps or emulators on their phones due to the limitations of the built-in calculator app.
• There is a need for a more advanced calculator app that can perform symbolic math and other complex calculations.
• The discussion around the phrase 'champing at the bit' highlights the evolution of language and the importance of understanding the origins of idioms.
• The behavior of NaNs in floating-point arithmetic is not always well-defined and can lead to unexpected results.
• The use of NaN boxing in some runtimes relies on the specific behavior of NaNs, which may not be guaranteed by the IEEE spec.
• The article's findings suggest that Apple's quality control may be lacking, allowing defective devices to reach consumers.